SPEAKER_00: two years ago, they weren't in the race. They weren't even in the stands. They weren't paying

SPEAKER_01: attention to this. They were changing the name of the company to Meta because of the Metaverse. Now, Meta means search engine for AI. He literally took the brand Meta and he's like,

SPEAKER_04: you know what? Meta.ai. It's Meta.ai. I meant it all along. Like literally, a guy fell on his face, got barbecued by everybody. Everybody said he's distracted and he's like got this stupid idea with the Metaverse that nobody cares about.

SPEAKER_09: And he just turns this entire judo move. This is a judo move. He redirected the energy. And I think the MMA stuff got him reinvigorated as an entrepreneur. He just took the battleship and

SPEAKER_12: turned it around. Congratulations to the team over there because you cannot f*** around with a guy that wears a chain on the outside of his shirt. I'm going to tell you this could be the most

SPEAKER_15: important news story of the year. This Week in Startups is brought to you by Open Phone. Create business phone numbers for you and your team that work through an app on your smartphone or desktop. Twist listeners can get an extra 20% off any plan for your first six months at openphone.com slash twist. Eppo. Experimentation is how generation-defining companies win. Accelerate your experimentation velocity with Eppo. Visit getepo.com slash twist. And Hidden Layer. Generative AI is revolutionizing industries. Hidden Layer's AI detection and response solution secures your generative AI and LLMs from malicious attack. Helping you generate more by enabling seamless and secure generative AI. Visit hiddenlayer.com slash twist to learn more. All right, everybody. Welcome back to

SPEAKER_23: This Week in Startups. And with me, after pulling an all-nighter, my guy, Sunny Sandeep Madra.

SPEAKER_24: How are you doing, brother? Big day. Big day. Llama 3. There's a lot to talk about. Lots to talk

SPEAKER_29: about. AI never sleeps. Just like you, Sunny, now that you are trying to get all of these new LLMs up and running on the Grok infrastructure. Everybody knows Definitive Intelligence. Your company was bought by Grok. Yeah. And now you are working with developers. And big news dropped this week from Facebook. Their open source model, which is called Llama, released version three. Explain to

SPEAKER_32: people why this is important and what it is. The best way to explain its importance. I'm going to pull

SPEAKER_35: up a chart from the, and you know, we've used this before, uh, is that it, um, is a new model. They trained, they open sourced it, but as people got their hands on it over the last 24 hours, that this open source model, 70 billion parameters. Okay. It's almost as good, just a little bit behind GPT-4, but better than Claude 3, which was at the top, better than Gemini Pro, better than Claude Sonnet, better than Command R, better than the original GPT-4. Wow. So here we go. The race is on. Yes. This

SPEAKER_23: is big news, big news. An open source model has arrived that's in essentially the top two models

SPEAKER_41: in the world. And it's more than an order of magnitude smaller. So not only has- Explain smaller

SPEAKER_26: in this context, right? A lot of folks are new to AI for sure. We're talking about parameters. We're talking about the versions of these models. We're talking about the context window. Let's do the

SPEAKER_00: audience a favor and just explain these from first principles. When you say size, do you mean the size of the software, the size of the package of the model? What does size mean in this context? So I'm,

SPEAKER_38: I'm going to, I'm going to use an analogy that I think makes it simpler for folks to understand,

SPEAKER_35: and then we can double click if you need to. The way to think about number of parameters is neurons.

SPEAKER_41: Okay. And the way to think about neurons is like in, you know, uh, animals, the more neurons,

SPEAKER_46: the smarter the animal. Got it. Like an octopus or a human or a dolphin or a whale,

SPEAKER_29: all those delicious sentient creatures that we eat. That's a shout out to this week's all in, which you haven't seen yet, but we had a little bit of a, a little bit of a touchy moment on our

SPEAKER_49: shared love of octopus, which occupy have a lot of neurons. Okay. So neurons, when we say

SPEAKER_41: neuron parameters, and the more parameters we have, the more neurons, the model has, the smarter it can theoretically be. Got it. And so the general thinking has been, we're going to create bigger and bigger models by bigger being more and more parameters. Got it. And the more parameters you have, the more training data you have to give it. And it's usually like several orders of magnitude, but let's just say one for simplicity. So if you have a 1 billion parameter model, you need to kind of give it 10 billion pieces of data, right? Okay.

SPEAKER_35: So if you have a trillion parameter model, you have to give it 10 trillion and then et cetera, et cetera, et cetera. What Facebook did meta did, they, they really changed things up. So one of their, one of their, you know, I guess moat says they have a lot of data. They have all the world's data.

SPEAKER_41: Yeah. They took a smaller model and kept giving it more and more and more data and what they've been able to show. And what's really interesting is Zuck in an interview yesterday said he, they stopped training this model. They had, it was still getting better and better. They needed to reallocate the GPUs to train Lama four. So they're like, we, we were blown away and it was getting better and better and better. So it's the same model architecture. Got it. Just given more and more data.

SPEAKER_64: And the data they have, yeah. Where did they get their data from? Have they been clear about this? Cause it's an open source model.

SPEAKER_13: Yeah. That means the code to build the software is open source. You can go see that code, but the open source data is different. So tell us, uh, what can we understand about the data in the

SPEAKER_41: model? If at all, anything that's not as clear, right. But you know, if you watch the interviews, one of the things that he said is they trained it with a lot more code and look, one of the things that they have, they have a lot of code. They've been running, uh, one of the world's largest services or several of the world's largest services for years. They have tons of code. And so they, they have every right to take that code and feed it if they want to, right. Um, they may have rights on,

SPEAKER_70: you know, I don't really know what Facebook's rights are on your, your eyes data, but I'm sure

SPEAKER_68: somewhere. I think they can use your data to train their model. That's yeah. You can be certain their terms of service has that basic, right? Yeah. Or, or they changed it recently or, you know, within the last couple of years to do it. So, um, you had to spend all night getting this up

SPEAKER_23: and running on the ground all day. Got it. Yeah. Yeah. Yeah. So this drops onto GitHub or hugging

SPEAKER_29: face and then just, you know, explain to the world here listening how that winds up being deployed

SPEAKER_79: on the grok infrastructure. Yeah. So for us, like, so these models are trained on NVIDIA hardware and when you're training a model, you're doing, and I'm going to do an oversimplification,

SPEAKER_41: you're doing a forward pass, which is like running the model. And then you're doing like a back propagation. And so you're in, in, in training them, you're running the models all the time. So once the model is done, you can run it on NVIDIA hardware. And so others can get it up and running. We obviously have our own custom hardware. So we have to take that model and we run it through, you know, our, we, part of our secret sauce is our compiler and we run it through that compiler. And then that compiler takes that model and basically runs it on the configuration of chips that we've given it. So there's two models that we're running, um, in the last 24 hours. One is the smaller one, which is Lama 3 8 billion. What's interesting is Lama 3 8 billion has the performance

SPEAKER_79: of Lama 2 70 billion. So it's got 10 times less parameters, but it has the same performance as

SPEAKER_81: one before fascinating. And then they, why is that? Why is that? Well, I think it comes back to,

SPEAKER_70: and look, everyone is being blown away by this, right? That's, you know, I was showing some tweets

SPEAKER_41: and I'll, I'll pull those back up in a second here. What everyone is realizing here is that maybe you can take, and you know, this sort of makes sense in biology. All of us humans have about the same number

SPEAKER_70: of neurons. Right. You know, I, I don't think there's like a 50% difference. Right. Right.

SPEAKER_35: If you give one human, a lot more data, their output capability can be a lot more than another one.

SPEAKER_41: Okay. So I think what we're learning here is that if you give more and more data to even a smaller model, the results that it can produce in these benchmarks are incredible. And I think that's, that's the path and, you know, Facebook again, was a lot different than any of the other companies because the core of their business has access to a huge amount of data is proving that they can build world-class models without even making them. Um, you know, GPT-4

SPEAKER_70: is generally considered to be a 1.7 trillion parameter model. So if you think about that, that's 20, more than 20 times larger than Lama 370B.

SPEAKER_88: Juggling multiple devices and apps to run your business is a mess. Open phone is here to make it simple by simplifying your business communications with one easy to use app. Open phone has rethought every detail of what a modern business phone should be. And here's the magic. It works through a beautiful, elegant app on your phone, or you can just use it on your desktop, making it super easy to get a business phone number for your entire team. And you know how brilliant open phone is. My teams use it every single day. My sales team loves it. My ops team, they use it all day long. And here's the features that we love. You can create a shared phone number, like customer support with multiple employees, fielding all the calls and all the texts to that one number. At my investment firm launch, we pride ourselves on replying to every single call or email instantly. And open phone is the number one rated business phone on G2 for customer satisfaction. So here's your call to action. Super easy. Open phone is already affordable. Starts at just 13 bucks a month. But twist listeners get an extra 20% off any plan for the first six months at open phone.com twist. And if you have existing numbers with other services, no problem. Open phone is going to port them over easy peasy lemon squeezy, no extra cost. Head over to open phone.com twist to start your free

SPEAKER_29: trial and get 20% off. Okay, so you've got the model up and running. And I was looking at an interface.

SPEAKER_93: It looks like meta is trying to let's play. Yeah, I think this is kind of important. If you go to meta.ai.

SPEAKER_95: Yeah, I got I got a cute up here for us. Okay, great. Because this is I started playing with this myself

SPEAKER_29: today in between meetings. Okay, so this is just on a very cursory basis. And I just want to make one observation when you pull up the meta AI's landing page. Yeah, it has a very unique style to it. Okay, fun

SPEAKER_00: style with a number of different ways to ask questions, but you don't have to log in. So I just want to pause on that. Okay, because not having to log in, there's another website where there's a search box and you can type something in that one doesn't require it either. So you can continue

SPEAKER_29: without logging in. And it just asks you what age you are because I think they've gotten a little trouble with age gating or whatever. And so they're being trying to be thoughtful of that you remember when suck was pulled in front of Congress or the Senate or somebody to discuss these kind of issues.

SPEAKER_64: But you know, I tried asking it who is Jason Calacanis and it did a wonderful job of like

SPEAKER_23: pulling from, you know, my Wikipedia page or whatever, a really tight summary. And it did seem to me like they've built an interface that looks very reminiscent of the Google original search page.

SPEAKER_70: I hadn't thought about it through that lens because when I see all these chat bots, they all kind of are starting to merge for me. But like now that you're bringing it up, it does have those vibes of like sort of an early Google very clean look search page. Yes. And you know, now we all use browsers. So I never even hit Google. We always just search out of our exactly the URL.

SPEAKER_26: And they told I saw Zuck in some clip of the interview he gave with his new haircut.

SPEAKER_29: We're mentioning the new haircut, I guess I guess he's got a new stylist or something. So somehow that his new look and his approach is interesting that people are noting, but he said they're going to put this search box on the top of every app they own. Yes. Okay, I want to pause for a second there. This is a shot across the bow of Google. And there's two reasons. The first reason is that they

SPEAKER_01: are both a place where people start their, you know, journey on the internet, right? You're on your phone,

SPEAKER_00: your open your browser right now, there's probably a tab open to, or an app open that's run by these two companies, whether it's YouTube, Google, Gmail, Facebook, Instagram, or WhatsApp. Okay, so they all, in other words, have massive distribution to billions of users. But there's never been a search engine

SPEAKER_04: to speak of, or a search function of the open web, um, in any of these products by meta. And now there is, and they're going to super distribute it. So if they can cut, if they can get one, two, three, four percent of searches away from Google, part two is they have an ad network already. They already have

SPEAKER_116: advertisers. And the advantage that, um, Google has always had is that in the box, you type your intent. Looking for a sushi restaurant, flying to Sydney. I'm going looking for a hotel in Dubai. There's a lot of purchase intent, right? In a search. Yeah. Oh yeah. Facebook's a window into the user's mind.

SPEAKER_00: Exactly. At that moment in time. And you can connect it with an advertiser. Facebook has had to connect

SPEAKER_123: based on psychographics. You're a 53 year old living in the Bay area. You're married, you have three

SPEAKER_29: kids. You like to go skiing. Your friends share pictures of you skiing. You also go to Mexico quite

SPEAKER_00: often. And we see a lot of sushi and, you know, oh, you, you, uh, seem to be, you know, uh, commenting

SPEAKER_124: on photos and following accounts in Dubai. We have some psychographic on you to give you ads. Well, now meta is going to have both direct intent and psychographics. This is going to create the greatest threat to Google search advertising dominance that's ever been created. And this is going to make meta

SPEAKER_89: stock and revenue go wild. Ooh, is the J trade is on. It's on, but is it, are you going to lever the J

SPEAKER_23: trade? Are you going to do it? Am I spreading? I'm going to spread the trade. Am I going to short

SPEAKER_127: Google long meta? Well, no, you can just lever it just by getting into call options. Cause that

SPEAKER_00: gives you some levers. I mean, listen, it's Friday night when we're taping this where, you know, the market is closed so we can discuss this and we got a little window, um, you know, before this drops, but I think I might go into my Robin hood account and I might J trade, I might sell half my positions and put it into meta. And now listen, I bought meta at $90. I don't know what's trading at now. 400, 500. It is trading $480. Okay. I've already run this bad boy up 5X. Okay. I'm 5X. And I'm thinking that this could, you know, be explosive. And the fact that Zuckerberg is talking about this in a direct video is extraordinary. It reminds me of the impact YouTube had on, uh, Google, Google had,

SPEAKER_29: you know, a certain type of advertiser, you know, a direct marketing type advertiser, right? You're, you're, you're kind of get like this little, uh, I refer to it as the sniper shot. Okay. You weren't getting, you know, like an emotional advertiser, psychographic advertiser. You weren't going after moms or dads or teens or single people versus married people. You get the idea when YouTube came along. Now you started,

SPEAKER_00: had the ability to show people video ads, which evoke emotion. It's like a whole different group

SPEAKER_136: of advertisers. And you combine those two things, you know, you see ads. Was it net new or was it a

SPEAKER_129: shift? Did like people move budgets or did the budgets just increase? This gave them access to a,

SPEAKER_00: a budget that they didn't have access to, which is some advertisers want to visually show their product. And so you take a car company. How many times have you seen car ads on YouTube all the time? How many times do you see a movie trailer or a TV show ad all the time? Why? Because you know,

SPEAKER_01: if you show a, a, a car, a Volvo, you know, a Toyota zip, zip, zipping along, you might buy it.

SPEAKER_00: And if you see, you know, Volvo is an interesting car in a, in a search ad, you're not gonna, it's not gonna evoke any emotion, right? So you have this emotional connection, this visceral connection that comes from video. That really, really, I believe was a key moment in the history of Google in terms of making a really great ad platform. So you had people who were marketers who now were, were able to spend two different ways.

SPEAKER_146: That's what's gonna happen here. That's what's gonna happen here with Facebook, you know,

SPEAKER_124: and I'm seeing it. It's like, you just see, you've seen this movie. I've seen this movie before, and I think this is gonna be explosive. I don't think necessarily

SPEAKER_00: Google's the loser here. I think anybody else with advertising revenue in the world,

SPEAKER_39: newspapers, televisions, you know, magazines, content websites. I mean, I just think it's

SPEAKER_150: gonna make them even more dominant. Correct me if I'm wrong, but it always feels like

SPEAKER_35: that from the earliest days, Facebook has sort of had the best relationship with their advertisers.

SPEAKER_00: And they've done pretty well. Yeah. Both of them. Google was Google had a very hands-off

SPEAKER_04: approach. They didn't, they wouldn't really talk to you that much, but YouTube did. Yeah. So, you know, those people who like to do video ads, that's like, let's go get a, you know, that's a madman, you know, let's get a three martini lunch. Yeah. Let's go to Cannes, you know,

SPEAKER_29: that whole group. And so, you know, I, I think you're right. Facebook did have like a little bit more of that, um, shake hands and, you know, go out for dinner kind of approach. And this is gonna just

SPEAKER_156: get them the direct response people in a really amazing way. Yeah. Wow. This is a game changer.

SPEAKER_52: The J trade's on. I think I might just go in there and just literally sell everything I own in the J trade portfolio and just put it into, you know, uh, Meta and Google. I'm doing these searches in

SPEAKER_156: real time while we're having a conversation. And let me just share my screen here because you've been to

SPEAKER_83: Dubai. Yeah. Yeah. So I'm going in the next couple of weeks again. Oh, okay. Well, maybe I have to

SPEAKER_134: jump on a plane with you because you know, I love that. Here we go. I could jump on the Grock, uh, Gulfstream. No, no, no, no, no, no, no. Oh my God. Oh, that'd be so yummy. Yummy.

SPEAKER_29: Okay. So look here, four seat. Here we go. Here are some of the best hotels in Dubai, Four Seasons Resort, Dubai at Jumeirah beach. We know that one, the Palm, the Ritz one and only one, uh, Palm. We know that Bulgaria. We know, I mean, these are, these are the best and, um, you know, it gives citations and, you know, if you click on those

SPEAKER_39: citations, uh, one and two, uh, it's dumping you to a Google search. Oh, wow. I wonder if they're

SPEAKER_120: getting paid like the, um, the Apple arrangement. Isn't that interesting?

SPEAKER_88: Are you tired of slow A, B testing? I'm sure you are. Do you have any trouble trusting your experiment results? I know I do sometimes. Well, get ready to 10 X your experiment velocity with EPPO. That's E P P O. Whether you're a scrappy startup, a tech giant, or anybody in between their feature management platform will turn your risky launches into clear cut experiments. Data teams, of course, love EPPO. And so will your product growth and machine learning teams. The executive is going to love it too, because you're going to love the results and the discipline that comes from defining really important product experiments and then executing on them really well, because it gives data teams better coordination and faster innovation. Online marketplace Inventa has cut down experiment time by 60%. And ClickUp, the project management company, has cut down analyst time by over 12 hours. And EPPO's cloud-based system is all about empowerment. You get instant access to your experiments from anywhere in the world, boosting flexibility and teamwork. And you get a system that grows with your needs, easily scaling up to handle more tests as your business grows. With EPPO, the daunting becomes doable. I love that. So your team can rely on the experiment results and make faster decisions. Here's your call to action. Experimentation is how generation-defining companies win. Accelerate your experimentation velocity with EPPO. Visit getepo.com slash twist. Just visit geteppo.com slash twist. Now let's get some experiments running. Let's get that product market fit.

SPEAKER_170: And thanks to EPPO for supporting independent media like this week in startups and all the startups who are listening. Well done. You look at my screen here. Look at this. When I hit sources,

SPEAKER_124: it's another Google one. And it says the luxury, the two web pages where it took this information

SPEAKER_00: from. So that webpage now at Forbes is going to get some traffic. Okay. So, okay. This is interesting.

SPEAKER_01: Um, what is the, why is there a Google search down here? Does anybody know what's going on here? Is there a Google relationship we're unaware of? These are dogged competitors. So why would they link to a Google search or are they using the results of the Google search to get this information? And then number two, when you click view sources and it just has these links here, it puts the Google logo next to

SPEAKER_175: them. Why? So I can, I can speculate. So basically, you know, all, all these modern answer engines that

SPEAKER_35: are being built and, you know, we've, we did a demo before as well, these modern answer engines, right. Are using, they need search results and, uh, there's different services that have been created to do that.

SPEAKER_70: And so what, what has maybe happened here, you know, obviously meta is not going to just go use one of those services. They're not going to scrape Google. So they maybe just went and did a direct deal because

SPEAKER_41: what these engines need is the results to come back. And then the engines use their own reasoning

SPEAKER_70: capability to, to like, you know, read the results and then parse things out and then return it to you. That's a good answer though. Kokari is incredible. Look at this. So this is on search engine journal, which is actually, you know, uh, you're quite reliable exactly for, for, for this particular thing and say meta integrates Google and Bing search results into the assistant. Uh, and so that's what they're saying. Both are coming in and they're talking about that. And you know, they're talking about the Zuck interview where he brought this up. And so really interesting now. And here's the examples of bringing meta into all the applications. Amazing. Wow.

SPEAKER_29: And, uh, so this is going to just change everything. You know, Zuckerberg is now coming in and he's

SPEAKER_35: looking to steal, you know, I gotta, I gotta tell you, and I feel like you have one of these and you

SPEAKER_12: definitely had one back in the day, but when you break that chain out. Oh yeah. That means you're

SPEAKER_124: going gangster. This is, I mean, I suck. Zuck is going into his, um, you talking to me phase.

SPEAKER_00: You talking to me? I, I don't see anybody else here. So I think you must be talking to me. Zuck's hair is out of control is no gel, uh, or maybe he's putting clay in it. I don't know what's going on, but the hairs on fleek. He's wearing a gold chain around his neck. I think he's hanging

SPEAKER_134: out with these UFC guys. Yeah. His testosterone has gotta be off the charts. He's maybe he's doing human growth hormone. I don't know what's going on here, but this is a level of aggressiveness

SPEAKER_04: that is even like for Zuck a little bit out there, but yeah, this is a big deal, folks. This is like a weird, you know, weekend drop Thursday, Friday drop. And if you're at Google

SPEAKER_00: right now, if I'm Sergei, if I'm Larry, uh, if I'm Sundar, this is another war room moment. Now you've got Microsoft on one side with Bing and you got Facebook on the other side. And they're

SPEAKER_23: both looking at that search money printing machine and they're saying, yum, yum. Is there a chance to

SPEAKER_192: steal it? I saw this and I had to bring it up because I think it's like, it was good.

SPEAKER_70: Yeah. It was good. This is the three phases of Zuck. Obviously early Zuck from this might've been like a CNN interview when he was probably, you know, just, just, he just made it to San

David Friedberg: Francisco. He just left. This wasn't that long ago. This is the in front of Congress.

SPEAKER_197: Yeah. Well, he's been pulled up in front of Congress for the last five years. So who knows

SPEAKER_129: which one that was, is that the most recent one? The most recent one. Yeah. He's obviously someone did a little bit of touch up here and gave him, I think they gave him a goatee

SPEAKER_180: there. I don't think he's. No, he wasn't. No, no, he wasn't. But like everything else is pretty good, but this is pretty awesome. It's pretty awesome. Cause what they're saying is this is

SPEAKER_29: this famous bell curve chart. So you have like the elite person all the way on the right of the bell curve. You got the average mid person. And then the left, you got like the more raw person. So usually the way this is presented is like, there's a ninja hacker over on the right. There's like a low IQ. I don't want to say the term, but like mentally disabled person on the left. And then the middle is like everybody. Yeah. And what they, the point they kind of make is the people on either end of the spectrum are like the unique, interesting people in the world. And so here we go. I mean, kudos to Zuck, but the game is on now. And you know, this is one of the interesting things about antitrust by the

SPEAKER_04: time Lena con or, you know, the, the UK or the EU or any, uh, federal agency starts to take apart a monopoly, the free market takes them apart and the barbarians are at the gate. You got Microsoft

SPEAKER_00: coming in the front and now you've got on the side, boom, another person coming in on the side.

SPEAKER_180: One more thing from the interview that he did yesterday, or he did a couple, but the one with,

SPEAKER_35: um, dark wash, which he said something really interesting. Like, Hey, how did you get on top of the GPUs? He's like, well, we needed the GPUs to basically compete with reels. And we needed in order to compete or create real, sorry, that compete with tick tock. And in order to pull, in order to create the algorithms, they had to go from algorithms that were built for like,

SPEAKER_70: you know, for your own network to tick tock style algorithms, which is they have to find the most interesting thing in the entire network. And so they did a monster purchase in 2022. So it was really interesting. And then they ended up having all this compute, which they've been able to leverage

SPEAKER_41: and turn into all this wonderful stuff. Um, it's, it's really awesome. And, you know, what I'll say is the implication for the broader ecosystem. So once you look outside of that, and I'm just going to pull

SPEAKER_70: up another tweet. So we talked about what having, you know, more powerful models means, but if you really look at what's happened last 24 hours and why I left our poker game to come in and, and get this launched is we are running Lama three at 300 tokens a second. Claude Opus, which has moved down on this list runs at 18 tokens a second and GPT four is 36. So if you're building something and they're all in the relatively same bucket of, of capability, you want the fastest one. So that's what we've been

SPEAKER_26: very excited about over the last 24 hours. Okay. Yeah. And Reuters is reporting that meta has struck some kind of a deal with, uh, Google to include those results. And so something's going on here where the results are being provided. Um, and so we'll, we don't know the details of that, uh, deal,

SPEAKER_64: but it's, um, it's definitely a big deal because they're gonna, whoever gets that search result

SPEAKER_26: into their database is going to have a lot of power because you can connect that search forever to that

SPEAKER_04: person. So, you know, if you were to search for something like you're feeling sad and depressed now, you know, you're going to see depression, medicine, ads, pharmaceutical ads to the cows come home because you did a search, you know, like, uh, I'm how to feel less sad. Or if you're like, you know, you do an Olympic search, I want to lose weight, whatever it is, you know, whoever has the

SPEAKER_00: search, man, I mean, it's just such purchase intent. And you combine that with what pictures you're sharing and what's in those pictures. It's going to be incredible. So we don't have the details of

SPEAKER_12: what's going on here yet. We have two grades to give up two grades. Okay. We're going to give

SPEAKER_217: a one grade on the model Lama three and the second grade on meta dot AI, uh, a and a plus I'm going a

SPEAKER_29: and a plus. I mean, it's clear. This model has broken in with an open source model into the top two. Yes. And if, and if you want to say, Hey, one, two, and three, it could be wrong. It's a mistake. We need more time to figure it out. Even if it was in the top five, the top 10 models have been proprietary models, closed source models. Yes. Right. Yeah. So just the fact that it's even in the top three would be extraordinary, let alone, you know, number one or number two. Okay. So that's an A I'm not going to give an A plus. You don't have enough information, but next week we could reserve the right to raise our grades. Yep.

SPEAKER_134: And then for this interface, the reason I'm giving an A plus is not because the interface is there, but because the vision is there and it's not just the gold chain and the crazy hair and the fact that

SPEAKER_29: he's doing the product discussions himself. It's that he said in the product discussions that he's putting this at the top of every app and there's going to be a search box and he's going to super

SPEAKER_124: distribute it. That means he has made this a priority. War room micromanager. Like,

SPEAKER_79: yes, this is my company. I'm going to win.

SPEAKER_229: Listen, we all know generative AI is revolutionizing every industry. In fact, we talk about it here on this podcast every week and companies that are slow to leverage this technology. Well, they're going to risk falling behind their competition, but new innovations also bring new threats. How do you safely accelerate your company's gen AI adoption while also managing the risk? Well, let me tell you about Hidden Layer's AI detection and response platform. It's a security solution specifically designed for generative AI and large language models, LLMs. You know that. It provides your security team with visibility and tools necessary to detect and prevent threats, like leaking of confidential information. We've seen that a couple of times, as well as malicious prompt injection. That's a serious threat and model theft, right? People are

SPEAKER_231: starting to steal. That's just the truth of it, folks. People will steal your innovations.

SPEAKER_134: That's going to enable your company to harness the power of AI and to do that securely and

SPEAKER_229: confidently. It's trusted by leaders in finance, tech, healthcare, and even the US Department of Defense, the DoD. Hidden Layer helps you generate more by enabling seamless and secure generative AI. So here's your CTA, the old call to action. With Hidden Layer, go from pause to possibilities and step into the future of secure AI innovation. Learn how to protect your generative AI today by visiting hiddenlayer.com slash twist. That's hiddenlayer.com slash TWIST.

SPEAKER_04: You know that vibe, right? Somebody's here. He's not asking anybody in the company what they're

SPEAKER_00: going to do. He's telling them, put the search box at the top. We can discuss everybody's feelings about that, but we're doing it. And I tell you what, have a good time discussing it. I'm out. I got an MA. I got to get to Vegas to go see UFC 301 or whatever the hell it was that he was at. So this is, um, you know, uh, that's why I give it an A plus. If I was just judging the interface based on the searches I did, I give it a B or a B plus, but I think because the vision is there and, oh, I have one more thing to add. He created a look and feel and a specific domain name. And you can tell how committed people are on. If they give something a specific name and brand and let it stand on its own. Yeah. He's telling people go to meta.ai and I'll give you another reason. This is really important. He didn't say you have to log in. He's like, you know what? I know this is going to cost me money. Every time somebody searches, I know I'm going to be hitting the server from, I don't care. This is the investment I want to make. I don't know how much he's losing on these

SPEAKER_244: server farms, but he, you know, he was willing to lose 10, 20 billion dollars a year on silly goggles. He said a hundred. I know, I'm talking about the goggles. Remember he was losing 10, 20 million a year on the goggles that nobody was using 10, 10, 20 billion on goggles.

SPEAKER_00: Yeah. I mean, if he, if he gets, but 10% of search, it's going to change the entire industry. And so you give it a domain name, you give it a look and feel, you tell it's going to be super distributed. We're putting this on every single app. Yeah. I mean, it tells you everything you

SPEAKER_41: need to know he's going for it. And it's strong. Like, you know, everyone was kind of like, oh, meta now meta.ai. You're like, I could go there and I could search. I would use it.

SPEAKER_124: And you know, who doesn't do this? Google. Remember Google plus where did Google plus exist?

SPEAKER_29: Bard. Google bard. They never did bard.com. Google plus should have just done plus.com. Yeah. They never made its own look and feel Google plus. They were like, we're going to put

SPEAKER_254: it up in the little carrot. Like it was hard to get to even right. It was up in the top right hand

SPEAKER_29: of your navigation. It just showed that they were not committed to it as a standalone long-term brand. Whereas with YouTube, they were like, it's going to have its own office space. It's going to have its own logo. It's going to have its own vibes. It's going to have a creative space in San Francisco, a creative, I'm sorry, San Bruno, and then a creative space in LA. This is how, you know, when the CEO is committed is, do they put it in its own building? Do they give it its own brand?

SPEAKER_13: This is, I, I'm going to tell you, this could be the most important news story of the year.

SPEAKER_12: Wow. I think this might be the news story of the year for our industry.

SPEAKER_38: I am the other way around. I'm a plus on the model. Okay. And I'm a on the AI integrations.

SPEAKER_41: Okay. Say more. Well, the model, look, um, putting that out in the open and giving it to the community and giving them a challenge. And he also said this, right. Was I want the community to

SPEAKER_70: take it. I want them to make it better. I want them to, um, make it run faster. I want them to

SPEAKER_41: make it run cheaper. And he goes, even if from the community does something and it's 10% better, we're going to spend a hundred billion dollars on this stuff. That's a $10 billion saving for us. He also said, it's important that no one company owns the greatest technology ever invented. And so that's a really, really strong move in this world where most companies are trying to create closed AI and not just, you know, close the AI, but a bunch of others.

SPEAKER_29: Yeah. I mean, Claude's closed. Yep. You know, it's interesting also about that. I think when you're behind and you've got like a giant business, you're totally fine when you're behind being open source

SPEAKER_64: because it helps you catch up. So where was Meta a year ago or let's say two years ago, where was Meta

SPEAKER_70: two years ago in the AI race? No, two years ago, they were just buying the GPUs to try to compete with

SPEAKER_64: TikTok. Exactly. They were not in the race. They literally didn't have their sh they didn't have

SPEAKER_263: their running shoes on yet. Yeah. A year ago. They launched llama. They launched llama. So a year ago,

SPEAKER_01: they're in the race. And if there's 10 people in the race, but they were in ninth or 10th place.

SPEAKER_134: Yeah. Trailing the pack. And now you're telling me, and I trust your judgment, Sandeep. Yeah. You're telling me they're neck and neck at the front of the pack. Well, it's not just my judgment. I'm just

SPEAKER_120: sharing what the, the, the standard industry benchmarks are saying right here. Like they are tied for two.

SPEAKER_07: Yeah, exactly. So let's just pause again and recap what we've learned. Two years ago,

SPEAKER_00: they weren't in the race. They weren't even in the stands. They weren't paying attention to this. They were trying to, they were changing the name of the company to Meta because of the metaverse.

SPEAKER_01: Now Meta means search engine for AI. He literally took the brand Meta and say, you know what?

SPEAKER_04: Meta.ai. It's Meta.ai. I meant it all along. Like literally he fell on his face, got barbecued by everybody. Everybody said he's distracted. And he's like, got this stupid idea with the metaverse that nobody cares about. And he just turns this entire

SPEAKER_00: area. Judo move. This, I think that the, this is a judo move. He redirected the energy. And I think

SPEAKER_09: the MMA stuff is got him reinvigorated as an entrepreneur. He just took the battleship and

SPEAKER_12: turned it around. Congratulations to the team over there because you cannot around with a guy that wears a chain on the outside of his shirt. I mean, you might have to break one. I'm breaking it

SPEAKER_124: out. I got it right here in the draw. I'm breaking it out. The point is this chain is indicative of where he's at, which is you talking to me, Zach. This is just incredible. It's incredible. I, you know,

SPEAKER_52: and like I said, if he can get one or two or three percentage points of search, there's no reason he

SPEAKER_00: can't beat being at search. So now you gotta be Microsoft going, oh man, this lunatic, but open source allows you to catch up quick. And then he can take this high road position, like Elon has taken, which is, Hey, participate in my participate in my ecosystem. Don't use their software. Yes. And, uh, you know what he would never, he, he specifically turned off the app ecosystem inside of Facebook. I remember he screwed every developer who ever, every partner, whoever worked with him, he screwed from his personal partners in the business, uh, to, uh, advertisers, to people building pages and content providers, content providers, app providers, and his own co-founders. He screwed everybody. Now he's going, you know what, let's build together. Smart move.

SPEAKER_35: I don't know if I trust them, but we'll see. I mean, look, I think there is so much going to happen. I'm going to put another prediction out there. Jake. Here we go. Okay. The next 60 days.

SPEAKER_70: This is really going to shake up. I mean, if you're any of these other companies, right? You're really sitting there and saying there is a open source free model that's available. The only people that really can't use it are like, uh, other internet companies, right? Their license limits you if you have. Oh yeah. Yeah. I think the license says something like if you have something more than like a hundred million active users, you can't use this, which it's just targeted at like Snapchat

SPEAKER_279: and Tik Tok or, you know, really, that's interesting. Then how can it be open source? Then I guess it's, I've never heard of it. I've never heard of an open source license.

SPEAKER_86: Open source have different licenses, right? But have you ever heard of a license

SPEAKER_26: based on the number of users? That's the first time I've ever heard of that. I've heard of licenses

SPEAKER_29: like you have to contribute back. Yeah. You have to give credit link back. So anyway, this is, I think somebody's got to get in there and read that license bullet by bullet point, uh, because this does matter. And if you're going to build on somebody else's system, you really do

SPEAKER_153: want to understand that. But yeah, I, I agree with you. This is, um, it's another one of these like

SPEAKER_12: earth shattering moments. I'm just going to pull this up, Jacob. Cause this license is actually pretty

SPEAKER_70: simple and clean. You want, I want to give them credit for it. So obviously here's a license, plain English license. Obviously there's an agreement. There's a licensee that's you met a llama materials and then defining the right. Exactly. They grant your right redistribution. Here's the

SPEAKER_41: only thing that you need to understand additional commercial term. I'm just going to read this for the folks that are listening. If on the meta three llama version release date, the monthly active users of the products or services made available by or for the licensee or licensees affiliates is

SPEAKER_70: greater than 700 million active users. Then you must request a license from meta and meta may grant it to you at their discretion. So this is a very specifically targeted at like two or three companies because this doesn't affect anyone else. It doesn't affect the corporate like an enterprise.

SPEAKER_29: It doesn't affect anyone else. Yeah. I mean, it's rarefied air to get to that number. They're probably including, I think Reddit probably breaks that number of monthly active views. I think Reddit might be right there. I wonder if that's a Reddit. 700 million. Twitter's at, um, monthly, um, we're doing monthly active users. We got to get the monthly active users. Twitter's in like, like the,

SPEAKER_297: like I think a hundred or 250 years, like that kind of range. Yeah. Top monthly active users,

SPEAKER_192: 20, 24. Yeah. 700 million is a big number. I mean, it's like a top internet services by now.

SPEAKER_300: I'm just going to ask meta. Yeah, it's meta. That's it. That's meta AI right here.

SPEAKER_180: Okay. Facebook, 3 billion WhatsApp, 2.78. 2.5. Instagram, 2 billion. We chat, tick tock, telegram. Then you fall away off. There you go. But look who's at 750. Snapchat.

SPEAKER_303: You're right. It is a Snapchat rule. This is a Snapchat rule. Exactly.

SPEAKER_141: Or TikTok and tick tock right here. Yeah. I don't think

SPEAKER_300: tick tock's got their own jam. I don't think they need a monthly mouse Reddit 2024. Let's see what

SPEAKER_306: Reddit has. Dude, this is so good. Like I'm just using it here. I'm such an idiot. I'm like

SPEAKER_244: literally using a Google search. We're talking about this, replacing Google. This is why habits

SPEAKER_307: take a time. I know monthly active users for Reddit and Twitter. Oh, interesting. It said 1.2.

SPEAKER_01: Wow. That's interesting. I think because of all the SEO traffic they get. Yeah. Okay.

SPEAKER_00: I bet it had mine get a different answer. I'm getting 400 million in 2022. Twitter had 440 where

SPEAKER_04: it had 430 in 2020. Yeah. Okay. So anyway, it's I think probably Twitter and Reddit are bouncing up against this and obviously Snapchat's over it. That's a very interesting clause. That's the first

SPEAKER_70: time I've ever seen it. Oh yeah. That's been around from the earliest licenses. So that's not a new thing.

SPEAKER_124: Okay. Hmm. But you could still look at their code base and be inspired by it.

SPEAKER_191: Oh, look at this. Look at this. Twitter, 550. Yeah. It's growing. Yeah. Uh, yeah. It's yeah.

SPEAKER_124: All right. Well, there you have it folks. Is there anything else we want to demo or talk about when it

SPEAKER_79: comes to this crazy meta moment? I'll just do like one more kind of fun demo just for, for the sake of

SPEAKER_70: it, because it's really fast and it's super exciting. So one of the things that they've done is, uh, made it so that it can code incredibly well, I'm going to pull up an example here.

SPEAKER_322: Give me a second. So we were trying this last year in you, right? Uh, me, a snake.

SPEAKER_323: And you can see, see how fast that is. Yeah. And so, and so basically it gives you two files, uh, uh, right here's, you know, the, the script and, uh, the index on HTML.

SPEAKER_35: And if you bear with me for one second, I'm just going to pull that up. I'm going to save those

SPEAKER_192: and I'm going to pull them up and you're going to, you're going to get a, you're going to laugh.

SPEAKER_70: They're going to get a kick out of this one. This is where the internet has, has come to. So I just saved those two files because you can see this. I got a snake game.

David Friedberg: Yeah. You made the snake game. Yeah. Yeah. Perfect.

SPEAKER_290: I made it in like three seconds.

David Friedberg: Incredible. Yeah. Somewhere Atari, 80 year old Atari engineers are banging their heads on the wall.

SPEAKER_188: Yes. This represents like their careers. Like literally that was represents like was is like five years of Steve Wozniak's life. And it's like, yeah, Nolan, Nolan Bushnell just jumped off a roof. He's like, what?

SPEAKER_29: I mean, this does also, I think, lead us to the discussion we've had many times, which is an organization as big as Facebook with as many developers it has, as it has using copilots and building a at the same time is like the snake eating its tail. They are, because they're building the LLM that's going to help them be better coders. They're going to use that LLM to write better LLMs

SPEAKER_00: and code. And so this is when we talk about the pace increasing and people talk about, we're going to reach AGI artificial general intelligence faster than people think. When you see, I think this is another proof point or evidence that the velocity is increasing. When you see new entrants take the number one or number two slot or the number three slot, and it starts changing like this, what's happening is, you know, these, these models and these tools are making the people build the tools and the models better every day, every week, every month. You must be seeing that up close and personal as well. Maybe you could talk about what's happening in the developer community in terms of

SPEAKER_70: people getting better at building product. Yeah. So there's, there's an interesting framework here

SPEAKER_35: that I'm going to share with you that I think you should think about for your cohorts of companies. When human or humanity went through the industrial revolution, I'm going to use two examples. We went from bespoke car making, which means like a team would make like one car a day, like a factory where

SPEAKER_70: you got a thousand a day and you went from like farming for maybe your family and your, your village at most to like industrial farms. You can feed entire states and or countries. Yes. Right. So that's like the industrial, you know, revolution impacting our society. We really haven't seen this in technology because if you have a designer on staff and that designer, you ask to create a mock-up for you, right. Or if you have an engineer on staff, you ask the engineer to build something for you. What we now have is the industrial revolution for those folks, because I can now have a thousand images generated in just a few minutes. Yeah. Right. Or I can have 50 code games generated

SPEAKER_41: and one of the things that suck did talk about also. So that's like the first thing. The second thing is you can now use these models going, feed them, create data to feed themselves, to get even better.

SPEAKER_35: That moment is here for us now. And so the folks that are using these tools are starting to accelerate and go faster and faster and faster. And my guess is that more so than anyone else, they've really cracked that. And they've really figured out how to put AI itself in their loop of driving AI forward.

SPEAKER_29: Part of the industrial revolution was, you know, there were different machines. And to your point, we did have this happen in the information industry. It was called the printing press. Now, if you wanted to have knowledge, you had a priest, a monk sit there and they would take a book and then they would take some parchment and they would write a copy of the book. They would try it. And then a month later, they would have rewritten some text and then they would sell it to some prince or king and they would put it in their library. And then a printing press happened and they would make a metal template of that book. Set that knowledge or book. And they would set,

SPEAKER_00: you know, the knowledge there. And then they would press a button and it would stamp the paper and cut it and put it into a book. And then all of a sudden everybody could have that book, the printing

SPEAKER_29: press, right? The printing press changed the world. And then you also had that happen with looms and

SPEAKER_00: building of clothing, um, where you're, you remember these looms that people started using to make clothes and all of a sudden somebody would be making a sweater and it would take them a week

SPEAKER_29: and now it would take them a day. And then a machine eventually wound up doing it in, in minutes.

SPEAKER_00: And so that is, I think a good way to look at it. If an app took a person, if, if building an app as

SPEAKER_04: a developer, um, and a designer was the equivalent of that monk rewriting the book and it took them a

SPEAKER_32: month or let's say it took them three months to rewrite the Bible. Okay. Now they could make a hundred

SPEAKER_00: Bibles in a day. Could you make a hundred apps in a day with one developer? And the answer is yes.

SPEAKER_124: Yes. Yes. It's coming. It's wild. It's well, it's here, but it's not polished enough.

SPEAKER_29: And the answers aren't correct. And this is part of why the Hume AI pin, I don't know if you saw uh, Mark has, you know, barbecue and says it's the worst product ever. Mark has said it's the worst product he ever reviewed. But did you watch the actual review? I haven't seen it yet. I watched

SPEAKER_342: those with my son. So we're going to watch it tonight. So anyway, you watch it. What you'll

SPEAKER_29: find is it's a love letter to the hardware device. It's he's incredible about, you know, the build quality and how unique it is. There are some issues with battery life being short and it's a little bit heavy and it gets hot, you know, things you could normally see. And he has, but the problem he has with it is the price of it. The fact that it's not as good as his phone and it takes 12 seconds for the LLM to give him the wrong answer. So all the hardware he thinks is incredible, but the LLM is wrong. And so we just need to make sure the answers coming out of these LLMs are correct because you know, people were like, oh, he's beating up on these things. Well, the truth is, we haven't been as critical of these LLMs and the answers they're giving.

SPEAKER_41: But Jacob, we just saw how they're getting better. They're not using their training data. In those two examples, it went out to the internet and then it reasoned over the answers. Yes. That's the future, J Cal, not, you don't want it to recall from its training data. Yeah. That's when you're subject to hallucinations.

SPEAKER_70: Yes. But if you tell it just the same way you or I would, Hey, go and get this information, parse it for me and give it back to me in a table. That's what it just did there on those results.

SPEAKER_39: Yes. And it was much better. Yes. And it has a citation. So, and now, uh, since it's giving a

SPEAKER_04: citation that leads to our ongoing discussion about rights, if you give citations. And I think we

SPEAKER_29: started when we started doing these, uh, you know, weekly AI discussions, I said, if you give citations and you link back, I have a lot less problem with this because, and then I don't know if you saw the,

SPEAKER_369: uh, Adam Schiff, um, or, uh, yeah, we talked about it a little bit, right? We talked about a little

SPEAKER_29: bit. Yeah. How they were, you know, there's, um, agreement that you just have to tell it what you

SPEAKER_124: trained it on. I think Zuck is realizes that that's how this will shake out. And he's like, you know what? I'm just going to link to the pages that Google has in the top results. Zip, zip, zip. I did a deal with those guys. I did a deal with those guys. And I'll do a deal with you. And if you don't want to be in the index, just put no robots at TXT and we'll exclude you from these

SPEAKER_29: results. Yeah. But man, this is disruptive. This is disruptive. Uh, another crazy week. Congratulations, Sandeep on getting this going. If you want to go to console.grok.com, is that the right place for people to go to start playing? That's the way they can go and they get, well,

SPEAKER_180: they go to grok.com and just play with it, but you go to console, get a key and start building your

SPEAKER_29: own stuff, which is what we want you to do. And, uh, he's Sandeep on the X platform, X.com slash Sandeep. He's super active, getting the replies there. Say hi to him. Tell him you

SPEAKER_123: watch the show. Have a great weekend, everybody. The world is moving out of breath. The three,

SPEAKER_00: two, the world is moving at an ever increasing velocity. And the way you're going to get through that is to be a part of these discussions we have here every week. Tell your friends about this week in startups, search on YouTube for this week in startups, hit subscribe, hit the bell, make sure you don't miss an episode so you can stay informed. And so you can figure this all out with us. Uh, Sonny, any, uh, thing people should know about in terms of, I don't know if you have a startup program or a way for people to learn more. We gave them the URL, but is there a startup program there

SPEAKER_136: at grok yet? Or we do just hit me up. We'll sort you out. Okay. Let's just make a, uh, you should

SPEAKER_04: make a URL slash twist. I'll give you a free maker. Let's do it. If you make, if you make grok.com

SPEAKER_29: slash twist, that's twist. Yeah. Yeah. For my advertisers here, since you do the show and you

SPEAKER_04: put so much work in there, give you the free ad, you send people to that, it'd be the this week in startups deal. And you can just offer people like some deal and, you know, webinars or whatever,

SPEAKER_156: but that's, that's my gift back to you and the grok team is, uh, rock.com slash twist. Is it

SPEAKER_199: GROQ or K? GROQ. Yeah. You're GROQ. You sure you're GROQ? Yeah. You're the Q. We have that

SPEAKER_04: trademark. You have the GROQ.com. I know that this is a little contentious. GROQ.com slash twist, T-W-I-S-T. At some point, we'll have a twist landing page with some offers. Uh, Sonny, thanks for sharing your knowledge. We'll see you all next time. Bye-bye.